OpenAI just hired the security firm that broke into it. On November 6, Hacktron — the researchers who chained a community forum RCE with an SSO flaw to reach OpenAI’s internal GitHub repos — became its official red-team partner. The move is a blunt acknowledgment that AI integrations have created an attack surface that standard bug bounty programs can no longer cover alone.
The Exploit That Earned Them a Job Offer
On July 25, Hacktron researchers chained two vulnerabilities. First: a remote code execution flaw in the image-processing pathway of OpenAI’s Discourse-hosted community forum. Second: an OpenAI SSO weakness that let them pivot from that forum session into employee ChatGPT accounts. From there, they had access to internal GitHub repositories.
They didn’t read sensitive code. Instead, they opened a harmless pull request to prove access, then stopped and reported everything. OpenAI confirmed fixes in roughly 14 hours. The bounty paid out in September: $6,500 — covering only the OpenAI-side findings, since the Discourse vulnerability was technically out of scope.
That $6,500 would be worth $51,577 under OpenAI’s newly announced bounty rates. More on that in a moment.
What the Partnership Actually Tests
Hacktron’s specialty is the gap that opens when applications, identity systems, and infrastructure connect to each other. As OpenAI frames it, this is precisely “where traditional application security reviews stop covering” — the seams between services, the OAuth handoffs, the permission boundaries that exist on paper but are never verified end-to-end.
As AI agents connect more external services to OpenAI’s systems — Slack, GitHub, Linear, Notion, whatever your stack looks like — those connection points multiply. Hacktron builds AI agents for code review, penetration testing, and red teaming, and their approach is fast: one of their automated assessments ran 674 attacks against a hardened target in roughly three hours. That’s the kind of coverage manual review can’t match.
As part of the deal, Hacktron gets expanded access to OpenAI’s frontier models through Daybreak Red — a restricted tier for authorized security research that requires identity verification, usage monitoring, and legal attestations. Giving the firm better tools to find vulnerabilities is a calculated bet, not a gift.
The Bug Bounty Numbers Are Now Serious
Alongside the Hacktron announcement, OpenAI released updated bounty tiers on November 2. Critical vulnerabilities in ChatGPT, Codex, or the API Platform now pay up to $30,000 at baseline. Exceptional findings with quality and loyalty bonuses can reach $172,000.
The increases are steep:
- Account takeover findings: $6,500 → $51,577 (7.9x)
- High-severity findings: $1,000 → $3,162
- Medium-severity: $750 → $2,415
There’s also a bonus multiplier system — quality bonuses for well-documented proofs of concept, a five-tier loyalty program for repeat researchers, and time-limited priority campaign bonuses for newly launched features. Reports decided within the last 30 days are eligible for retroactive payment under the new rates.
If you do security research, OpenAI’s attack surface is now one of the better-paying targets available. The PROMPT2PWN event on November 20–21 in Oakland will put researchers directly against live OpenAI targets with the security team onsite.
What You Should Be Auditing Right Now
The focus of this partnership — third-party service connections — is a direct signal about where the real risk sits for anyone building on OpenAI’s APIs.
OAuth tokens are the specific problem. They operate outside SSO and MFA controls. A compromised token grants persistent access without triggering any of the authentication alerts your security tooling monitors. Industry data puts third-party integrations as the entry point in over 25% of API breaches, with 38% of data leaks traced to third-party API connections specifically.
Three things worth doing this week:
- Audit your integration graph. Map every service connected to your OpenAI API keys. Know what has access to what.
- Scope your keys. API keys should have the minimum permissions required. If a key can read and write everything, that’s a liability.
- Rotate tokens you haven’t rotated. OAuth tokens and API keys that haven’t been reviewed in six months are a risk. Treat them like passwords and rotate on a schedule.
The OpenAI-Hacktron partnership makes the threat model explicit: the attack surface isn’t your code, it’s the graph of services your code connects together. That’s what gets tested when skilled researchers go looking. If you haven’t mapped it, now is a good time to start — before someone else does it for you.
Related: Claude Code “Don’t Ask Again” Leaked 2,539 Secrets on GitHub: Fix Now | GPT-6 Cyber Found 81 Firefox Flaws and 948 Linux Exploits
